Guo Sikao, Korolija Nenad, Milfeld Kent, Jhaveri Adip, Sang Mankun, Ying Yue Moon, Johnson Margaret E
TC Jenkins Department of Biophysics, Johns Hopkins University, Baltimore, MD, 21218, USA.
University of Belgrade, Serbia.
bioRxiv. 2024 Dec 10:2024.12.06.627287. doi: 10.1101/2024.12.06.627287.
Particle-based reaction-diffusion models offer a high-resolution alternative to the continuum reaction-diffusion approach, capturing the discrete and volume-excluding nature of molecules undergoing stochastic dynamics. These methods are thus uniquely capable of simulating explicit self-assembly of particles into higher-order structures like filaments, spherical cages, or heterogeneous macromolecular complexes, which are ubiquitous across living systems and in materials design. The disadvantage of these high-resolution methods is their increased computational cost. Here we present a parallel implementation of the particle-based NERDSS software using the Message Passing Interface (MPI) and spatial domain decomposition, achieving close to linear scaling for up to 96 processors in the largest simulation systems. The scalability of parallel NERDSS is evaluated for bimolecular reactions in 3D and 2D, for self-assembly of trimeric and hexameric complexes, and for protein lattice assembly from 3D to 2D, with all parallel test cases producing accurate solutions. We demonstrate how parallel efficiency depends on the system size, the reaction network, and the limiting timescales of the system, showing optimal scaling only for smaller assemblies with slower timescales. The formation of very large assemblies represents a challenge in evaluating reaction updates across processors, and here we restrict assembly sizes to below the spatial decomposition size. We provide the parallel NERDSS code open source, with detailed documentation for developers and extension to other particle-based reaction-diffusion software.
基于粒子的反应扩散模型为连续反应扩散方法提供了一种高分辨率的替代方案，它能够捕捉经历随机动力学的分子的离散性和体积排斥性质。因此，这些方法特别能够模拟粒子自组装成丝状、球形笼状或异质大分子复合物等高阶结构，这些结构在生命系统和材料设计中普遍存在。这些高分辨率方法的缺点是计算成本增加。在这里，我们展示了使用消息传递接口（MPI）和空间域分解对基于粒子的NERDSS软件进行的并行实现，在最大的模拟系统中，对于多达96个处理器实现了接近线性的扩展。针对三维和二维中的双分子反应、三聚体和六聚体复合物的自组装以及从三维到二维的蛋白质晶格组装，评估了并行NERDSS的可扩展性，所有并行测试用例都产生了准确的解。我们展示了并行效率如何取决于系统大小、反应网络和系统的限制时间尺度，表明仅对于具有较慢时间尺度的较小组装体才具有最佳扩展性。在评估跨处理器的反应更新时，非常大的组装体的形成是一个挑战，在这里我们将组装体大小限制在空间分解大小以下。我们提供并行NERDSS代码的开源版本，为开发者提供详细文档，并可扩展到其他基于粒子的反应扩散软件。
